The AI Intervention Standard

Artificial-intelligence research often attributes an observed result to "the model" even when the result depends on data, objectives, training history, context, representation, inference policy, selection, internal intervention, deployment, and execution. The AI Intervention Standard develops a systematic framework for studying deliberate changes to configured AI systems. It defines nine locations of intervention, six evidence levels, ten principles, a twelve-rule research constitution, an eight-dimensional outcome vector, and a claim-specific Boundary Research Record. Repository evidence is used as a case base from which bounded principles are extracted rather than as the book's organising structure. The proposed standard connects measurement, counterfactual reasoning, preservation, delivery, lineage, and governance. It is intended for AI researchers, evaluators, red teams, model engineers, auditors, and research leaders who need to decide when an observed system change warrants a credible, transferable, and governable claim.

This publication proposes a research standard. It is not an accredited, formally ratified, legally binding, or independently certified standard. Its requirements and claims are offered for inspection, application, criticism, replication, and revision.